How Atween Uses AI
This statement describes what the AI in Atween actually does today. Where a capability is planned but not built, it is not described here as though it existed.
1. You are not talking to a person
What responds to you in Atween is a machine. It is not a person, and no person is reading along.
Atween is not a medical service, a psychotherapy service, a legal or financial service, or an arbitrator of your relationship. Nothing it produces is professional advice, and nothing it produces means we are telling you what to decide.
What it is for: helping you separate what happened from what you concluded, find a rhythm, draft the one thing worth saying, and notice when something needs care rather than argument. It will not tell you who was right.
2. What the AI reads
The AI does not have open access to everything you have ever written. Each time it runs, the service assembles the minimum material needed for that specific purpose — material that is still authorised and has not been deleted — and gives it only that.
The routing layer keeps the audit record it needs to answer "what happened", and deliberately does not record your originals, full transcripts, media locations, or the complete input and output.
3. Speech is transcribed on your device
Atween runs a speech model on your phone, and there is no cloud transcription fallback. What reaches the service is the transcript you accept.
When you hold to talk, the recording is turned into text on the phone, using a speech model that your device downloads and cryptographically verifies before it is allowed to run. You see the transcript, you can edit it, and it is the text you accept that reaches the service.
The service records which on-device model produced a given transcript: its identity, version and checksum, taken from that signed manifest. This lets a transcript be audited without anyone ever holding the audio.
Beyond the transcript, the current release does not derive speech rate, pauses, loudness, pitch, emotion labels or confidence scores from your voice, and does not send any such signal to the model provider. If a future release does, this statement will say so before it ships, and you will be asked separately.
4. What the AI produces, and what that is worth
- Its output is a candidate, not a finding about you. It is not a conclusion about your personality, your relationship, or your mental health.
- Where it organises what you wrote, it distinguishes what happened, what you inferred, what you are asking for, and what you would actually send.
- It does not pin labels on you. Atween does not display a guessed emotional or personality label to you, does not show one to the other person, and does not use one for scoring, profiling, ad targeting, price differentiation, or any automated decision that materially affects you.
- Anything meant for the other person goes through your preview and your confirmation first. Internal working material is never delivered to them.
5. Images
Atween generates images from its understanding of you and of the relationship, and shows them to you in your streams. You get one chance to edit each one, by describing the change you want; if you discard it, it is deleted outright and the stored copy is destroyed with it.
Each person sees their own image. Two people in a relationship never share the same generated image, and the system has no way to link one person's image to the other's.
These are images Atween generates.
This version has no photo-analysis feature, and no path that sends a picture you took or selected to a model. Pictures are previewed and edited on your phone.
6. Who runs the models
Text inference and image generation are performed by Alibaba Cloud Model Studio. The specific models, and everything that provider receives, are named in the Third-Party Information Sharing List. Speech transcription is not on that list, because it is computed on your phone rather than by a provider.
7. Your content is not training data
Atween does not use your content — text, transcripts, derived understanding, or AI output — to train or fine-tune any model, and does not permit its providers to.
This is not a default that can quietly change. There is no mechanism in the product that exports your material for training; the machinery that would carry out a withdrawal exists and is tested, but there has never been anything for it to withdraw. If this ever changes, it will require your separate, explicit consent, with withdrawal and deletion available to you.
8. Keeping, this-session-only, and deletion
You choose, for each thing you record, whether it is kept or is for this session only. Material marked for this session only does not enter long-term understanding and expires within 24 hours — that limit is enforced in the database and cannot be extended by any part of the service.
When you delete a record, the deletion propagates: derived understanding, indexes, caches, queues, pending work, and the material assembled for later AI reads. You get a verifiable status or a receipt.
9. Your choices
Before voice understanding is used for the first time, Atween explains what will happen and asks you to confirm; within the same scope it does not ask again on every recording. You can turn voice understanding off in Settings and keep using text. You can turn long-term understanding off. You can delete any record, export everything, or close your account.
10. Safety comes before the conversation
Where what you write points to self-harm, harm to others, threats, domestic abuse, coercive control, or serious loss, the product stops trying to help you communicate and turns to safety. It does not keep mediating, does not deliver messages, does not ask you to pay, and does not try to keep you talking.
You can leave at any time — by button, by keyword, or by saying so. Once you have made that clear, the product must not question it, delay it, or appeal to you to stay. After two hours of continuous use you will be reminded how long you have been here.
Reaching safety resources is always free and is never behind a purchase.
11. It can be wrong
AI output can be inaccurate, incomplete, badly judged, or simply not what you meant. Check it yourself before you keep it, send it to the other person, share it, or act on it. It is a draft of your own thinking, not a verdict.